Immune checkpoint blockade (ICB) therapy has revolutionized oncology, producing durable responses in some patients — but most do not respond. Predicting who will benefit from immunotherapy remains one of the most important unsolved problems in precision oncology, one that carries enormous clinical consequences and is notoriously difficult due to patient heterogeneity, tumor complexity, and the diversity of treatment regimens.
Existing biomarkers for immunotherapy response — tumor mutational burden (TMB), PD-L1 expression, microsatellite instability (MSI) — have limited predictive power in real-world settings. Most machine learning models trained on one cancer type or treatment fail to generalize to others. Clinical cohorts are small, follow-up is long, and the feature space is high-dimensional.
The core goal of COMPASS is to build a model that works across cancer types and across treatment classes — not just on the specific cohort it was trained on.
COMPASS integrates multi-modal clinical and genomic features — including tumor genomics, immune profiling, and clinical metadata — through a generalizable AI framework. Key design choices:
COMPASS achieves strong and consistent predictive performance across diverse cancer types and ICB treatments in independent validation cohorts — demonstrating true generalizability rather than cohort-specific overfitting. The model's predictions are clinically interpretable and can be computed from data routinely available in oncology practice.
A clinical-grade web server is freely available at immuno-compass.com, allowing oncologists and researchers to run predictions directly from patient data without any computational infrastructure.
This work was developed in collaboration with the Zitnik Lab at Harvard Medical School and Roche pRED, and is currently under minor revision at Nature Medicine.
COMPASS represents a step toward making AI-driven immunotherapy response prediction clinically actionable. By providing reliable predictions across cancer types, it can:
The lab continues to develop COMPASS extensions for new cancer types, combination therapies, and multi-omics integration under the AIDDPM research program.